Register for the iXsystems Community to get an ad-free experience and exclusive discounts in our eBay Store.

Sync performance issues

Not open for further replies.


Apr 17, 2017
I'm having a performance issue with my setup, When I have zfs set sync=disabled I get ~200MB/s sustained writes, When I do zfs set sync=always I seem to have a issue where the transfer starts at ~500MB/s (buffer?) for a couple of seconds then drops to about 200MB/s for the first ~6.5GB of transfer then it starts slowly decreasing the speed till it hits ~300KB/s then it jumps back up to 500MB/s and repeats the cycle again. I'm hoping someone might know the reason why it's doing this. It seems to act the same if I use the SLOG or not.

This is a test machine which is just a HP desktop with
FreeNAS 10.0.3
4x4TB desktop drives in a RAID-Z1
2x240GB Samsung EVO SSD for SLOG (I'm ordering better slog but this is just for testing)
1xWD Blue 250gb SSD for boot drive
Marvel pci-e SATA controller
32 GB ram
i7-2600 CPU
Intel X520-DA2 (10GB nic)

I also see the same sort of behavior on the NAS I'm setting up for production

and it has

FreeNAS 9.10.1
24x 2T HGST enterprise drives
2x Intel 320SSD for os
LSI 9211-8i controllers
64GB ram
2xXeon E5-2670 CPUS
(no SLOG yet)
Intel X520-DA2 (10GB nic)

Last edited by a moderator:


Neophyte Sage
Dec 11, 2015
Welcome to the forums.
Looks a tad wierd. I've no conclusive solution for you, but can offer a few comments.

1. Corral is sidelined, 9.10.2 is the relevant platform atm.
2. The speed "buffer" you mention is when data is cached in RAM, prior to being written out.
3. It doesn't seem like your SLOG works in that setup. I'd redo testing with a pool of 2 vdev mirrors. I suspect the underlying pool somewhat causes problems.
4. IIRC Marvel has been troubled with dodgy support on FreeNAS. That might be causing additional problems reducing performance of the SLOGS. Remember, latency > speed.
5. The production box - is it configured in pairs of mirrors?
6. Have you used any sorts of autotune?
Not open for further replies.